Typed two-head ANN router109 - source candidate

Typed two-head ANN router109 - source candidate

Status: SOURCE_ONLY_UNREVIEWED_UNEXECUTED. No module imports, compilation,
ANN inference, test execution, fitting, checkpoint writes or repository changes
were performed to author this package. Root review and a fresh exact execution
selection are prerequisites for running it.

This small component routes two explicitly selected authored tasks over an
already learned representation. It uses the actual persisted
learned-foundation-transfer104-v2 RESULT.v1 checkpoints for seeds11,29,47:
seed-N.learned-parent.json for origin-quadrant/v1 and
seed-N.head-on-learned.final.json for shifted-quadrant/v1.
The original checkpoint files remain unchanged. Both tasks see raw [x,y]
coordinates in dimensionless [-1,1], in that order. Source thresholds are [0,0];
target thresholds are [.2,-.2]. Action indices are neither, x-exceeded,
y-exceeded, both-exceeded. Same shape does not establish this semantic binding.

The four learned trunk tensors w1/b1/w2/b2 are exactly equal for each saved pair.
The source route retains its original policy wp/bp. The target route selects its
newly learned policy wp/bp. Target trunk array references are privately joined to
the verified source trunk; the original target checkpoint is not modified.
Each route retains its own saved untrained value wv/bv, which differs across the
two saved tasks. This value is reported as untrainedValue, never evidence of a
learned value/equity skill. No random or engineered feature substitute is created.

Interfaces

createSelectedRouter(seed) constructs from fixed inert embedded raw buffers.
createRouterFromCaptured(seed, sourceRaw, targetRaw) is the same admission path
for exact captured strings, not a path/provider/module loading interface. Only
seed11,29,47 and their exact two recorded identities are admitted. Changes even
to whitespace refuse the original file identity before parsing.

The instance exposes bindingFor(task), infer(request) and status().
A request is exactly {task, binding, observation:{x,y}}. binding must equal
that router's full frozen bindingFor(task), including seed, complete original
checkpoint binding, selected content/file IDs, encoder/action/geometry and
actual learned feature-basis digest. The full logical source/target task IDs are
mandatory. There is no automatic/latest task selection or shape-based coercion.
The coordinate signed zero accepted by the original encoder remains supported.

infer validates the request and binding before its sole numerical dispatch.
Its immutable output includes probabilities, actionIndex/label, the untrained
value, selected checkpoint/basis identities, optimizerUpdates:0 and
noAuthority:true. No proposal executes an effect. Status counts successful
router forwards, not a performance/quality metric. Calls4097+ refuse on that
router; there is no implicit lifetime reset.

Fixed source graph and identity

Production files are router.mjs, inference.mjs, selection.mjs, checkpoints.mjs.
Only Node crypto/buffer are imported in production. No fs, original harness,
registry108, training module, callback or model loader is imported. The original
harness reads its module sources during bootstrap, so that graph was deliberately
not imported without review. inference.mjs is the exact original numerical
forward body (lines53..84), with a fixed2/4 enclosing configuration and export.
Operation order is preserved; no optimizer/gradient/initializer path is present.
It is an internal pure numerical helper, not a model-admission authority API.

Original raw pretty-JSON SHA/byte identity is distinct from the harness's
sorted-key canonical checkpoint content digest. The saved fit receipt selects
content IDs; router admission recomputes them in JS. Full binding/architecture,
team/model grammar and actual trunk compatibility are checked separately.
STATIC-CHECKPOINT-IDENTITIES.json records generic lexical verification using
saved JS numeric tokens, sorted ASCII keys/strings and no float recomputation.
This static check is not an ANN inference or a JS module execution.
Selection sourceSetSHA256 is the actual frozen SOURCE-FILES.json digest.
Original interface-source and module hashes remain historical binding facts;
they do not certify this new source or authenticate caller identities.

Proposed validation and limits

router.test.mjs declares22 meaningful source fixtures, all UNEXECUTED.
It includes malicious source/target relabelling with identical shapes, changed
trunk/head/version/raw identity, wrong encoder/action/basis/seed, malformed and
nonfinite inputs, getter refusal, signed zero, separate scalar inverse checks,
interleaved source/target calls and immutable outputs. A full-grid fixture uses
the actual saved100 source-audit and100 target-evaluation rows for all3seeds and
compares reported original loss/accuracy within1e-12. This proposes a narrow
routing/retention witness, not a newly executed gate or independent discovery.
The prospective full dispatch is4709 router forwards +6 scalar-reference
forwards, zero optimizer updates. Exact declarations/counts are in the inventory.
A future root-selected command may use node --test router.test.mjs; this text
is not selection or permission to execute now.

New prototype limits are16384 raw string code units,65536 captured canonical
UTF8 bytes,8192 capture nodes, depth32, arrays4096, objects256keys and4096
successful inference calls per router. All selected saved raw files fit8611B
maximum. These limits do not change any original ANN budget or Registry99 cap.
Source-cost and native/loader/heap/time/RSS liabilities are explicit in
SOURCE-COSTS.json. The JS parser/canonicalization may allocate before the final
byte-size check. Proxy/intrinsic tampering, huge foreign objects, hostile process
security and arbitrary code effects are outside this ordinary local-data
premise; own descriptor checks alone are not a security boundary.

Remaining work

Independent exact-source review, any necessary source correction, a fresh
bounded inference-only execution selection and actual independent readback are
still required. This code does not qualify useful/general transfer, statistically
independent retention, hidden-feature retraining, optimizer resume, generic
registry adoption, poker/gameplay, containment or original W0/ANN resource fit.
All older failures and SPENT epochs remain unchanged. The full original ANN,
game/security and retention objectives remain open in parallel.

The author previously wrote/reviewed registry/codec/portable provenance and
performed persisted-outcome readback; those overlaps do not supply an
independent correctness verdict for this source. No managed worktree or lease
was allocated. HANDOFF / RETAINED_RELEASED when frozen.